For the Total Recall hackathon, developer Bigmanmarsh unveiled Suisub, a Telegram-based money agent designed to bridge the gap between passive observation and autonomous action in personal finance. Unlike traditional bots that reset after every session, Suisub persists its behavioral models using Walrus Memory on Mainnet, allowing it to learn user habits over time. The agent integrates with the Sui blockchain to detect incoming USDC payments and execute micro-transactions for data plans and savings, effectively acting as a persistent financial companion that lives entirely within a chat interface.

Persistent Memory and Trust Building

The core innovation of Suisub lies in its ability to rebuild its behavior from stored memories every time it starts. The developer demonstrated that when moving the agent from a local laptop to a free server with no saved files, the bot automatically restored all 12 memory entries from Walrus Mainnet. These memories track specific user preferences, such as savings rates, network providers (MTN, Airtel, Glo, or 9mobile), and data usage patterns. The system employs a gradual trust model: after a user approves a specific action, such as saving 50% of incoming funds, three times in a row, the agent offers to perform that action autonomously within set limits. Users retain full control via a simple /stop command, which instantly revokes autonomous permissions.

Integration with Tatum and Machine Payments Protocol

Suisub utilizes Tatum’s gateway to read Sui Mainnet, monitoring for incoming USDC and checking balances. When the agent executes a purchase, such as buying a data bundle, it interacts with the shop’s server, which responds with a '402: pay first' status. The agent then uses the Machine Payments Protocol to pay in USDC on Sui before retrying the transaction. This workflow eliminates the need for traditional card checkouts, streamlining the payment process for micro-transactions. However, the developer noted that while the USDC payment is real and on Mainnet, the actual delivery of the data bundle is currently simulated because VTpass API access was not approved in time for the hackathon deadline.

Real-World Testing and Honest Failures

The project’s documentation highlights a critical lesson learned during real-money testing: a configuration error resulted in 2.34 USDC being sent to a dead test address. This incident prompted the addition of transparency features, ensuring the bot now announces savings and payment destinations every time it starts and displays them in the /status command. Additionally, the developer discovered that Tatum’s Mainnet node retains only recent history, causing the agent’s initial payment detection method to fail silently. The solution involved adjusting the search parameters to start from a recent point and adding warnings for failed checks. The agent’s learning curve is visible in its data estimates; after initial inputs of 5GB monthly usage, subsequent 'I ran out early' feedback raised the estimate to 7.8GB, dynamically adjusting future suggestions.

Key Takeaways

  • Suisub uses Walrus Memory to persist agent state and behavioral rules across different hosting environments without local file storage.
  • The agent employs a trust-based automation model, requiring three consecutive approvals before offering autonomous execution of tasks like savings.
  • Payments are executed via the Machine Payments Protocol on Sui Mainnet, though bundle delivery remains simulated due to API limitations.
  • Real-world testing revealed critical infrastructure issues, such as silent failures from Tatum’s history retention limits and human error in address configuration.
  • The project is open source, allowing developers to inspect how memory-driven agents can manage personal finance workflows in a chat interface.

The Bottom Line

Suisub proves that persistent memory is the missing link for trustworthy autonomous finance, turning a simple Telegram bot into a proactive agent that learns from its own failures and earns user confidence one transaction at a time.